DISCLAIMER: This document does not meet current format guidelines Graduate School at the The University of Texas at Austin. of the It has been published for informational use only. The Report Committee for Matthew David Bramble Certifies that this is the approved version of the following Report: Feature-Based Clustering of Stomach Cancer Gene Expression Data APPROVED BY SUPERVISING COMMITTEE: Peter Mueller, Supervisor Christopher S. Sullivan Feature-Based Clustering of Stomach Cancer Gene Expression Data by Matthew David Bramble Report Presented to the Faculty of the Graduate School of The University of Texas at Austin in Partial Fulfillment of the Requirements for the Degree of Master of Science in Statistics The University of Texas at Austin May, 2018 Dedication Julie, thank you for your encouragement and support. Abstract Feature-Based Clustering of Stomach Cancer Gene Expression Data Matthew David Bramble, M.S.Stat. The University of Texas at Austin, 2018 Supervisor: Peter Mueller This report presents the results of using a probabilistic clustering technique in the analysis of microRNAseq and RNAseq data from gastric cancer tumor samples deposited at TCGA (The Cancer Genome Atlas). Using the method of Hoff, who has proposed a Dirichlet process unsupervised clustering framework with feature selection, it is possible to reveal interesting structure in gastric cancer gene expression data that relates to Epstein- Barr virus (EBV) microRNA levels. This structure is not as readily identified by a typical hierarchical clustering method, and the results of this analysis contribute to an understanding of the role of EBV viral microRNAs in gastric cancer tumors. iv Table of Contents List of Tables ..................................................................................................................... vi List of Figures ................................................................................................................... vii 1. Introduction ......................................................................................................................1 1.1 Background ...........................................................................................................1 1.2 Target Identification .............................................................................................2 2. Clustering With Feature Selection ...................................................................................5 2.1 Setup .....................................................................................................................5 2.2 Model ....................................................................................................................6 2.3 Posterior MCMC...................................................................................................7 3. Data Analysis .................................................................................................................11 3.1 Data .....................................................................................................................11 3.2 Clustering With Feature Selection ......................................................................16 3.3 Analysis of Relevance ........................................................................................23 4. Conclusion .....................................................................................................................30 References ..........................................................................................................................32 v List of Tables Table 1. Comparison of Clusters Under Different Orderings ............................................18 Table 2. T Test Results of EBV+ vs. EBV- Tumors and EBV+ High-Load and Low- Load Tumors ..........................................................................................................29 vi List of Figures Fig. 1. Distribution of EBV microRNA Loads Fig. 2. Expanded Region From Fig. 1 ..13 Fig. 2 Distributions for Curated Gene Features .................................................................15 Fig. 3. Complete Cluster Assignments ..............................................................................17 Fig. 4 Correlation Between Human and EBV microRNA Levels .....................................19 Fig. 5 Hierarchical Clustering of Gastric Cancer Data ......................................................22 Fig. 6. Results of Testing Different Beta Prior Parameters ...............................................24 Fig. 7 Partial Cibersort Output ...........................................................................................28 vii 1. Introduction 1.1 BACKGROUND The National Cancer Institute’s estimate for new gastric cancer diagnoses in the U.S. for 2018 is 26,240 patients, and the estimated number of deaths from gastric cancer for 2018 is 10,800. Only 31% of gastric cancer patients will survive past 5 years.1 The rate of occurrence of Epstein-Barr virus (EBV) association among all gastric cancers is roughly 10% worldwide, ranging from about 17% in the US to 4-5% in China.2 EBV is a ubiquitous herpes virus that is transmitted orally, establishes a lifelong latent infection, and is thought to play a role in the oncogenesis and development of gastric cancer. EBV also is one of a small number human viruses to express microRNAs (44 identified mature microRNA transcripts) that act analogous to human host microRNAs.3 Herpes viruses account for most known viral microRNAs, and these microRNAs aid the viruses in maintaining their hallmark latency through increasing the longevity of infected cells and allowing immune response evasion, among other effects.4 With respect to the genesis or development of epithelial cancers such as gastric adenocarcinoma, additional effects relate to cell proliferation, transformation, and other wide ranging effects that aid in the development and persistence of tumors.5,6,7 Potential effects also, of course, include the broad range of cancer-related effects that have been found in relation to human microRNAs.4,8 MicroRNAs, both human and viral, are typically short RNA sequences of roughly 22 nucleotides that bind with the RNA induced silencing complex (RISC) in the cytoplasm, thereby allowing regulation of messenger RNA levels through complementary binding, typically in the 3’ untranslated region of messenger RNA transcripts. In humans, the canonical microRNA processing system involves transcription of a primary microRNA 1 sequence of hundreds to thousands of nucleotides in length, processing by the RNase III enzyme Drosha and the dsRNA binding protein DGCR8 in the nucleus to produce a stem- loop structure of about 70 base pairs, transfer to the cytoplasm via the nuclear export factor Exportin 5, and processing in the cytoplasm by a second RNase III enzyme, Dicer, to create two 22-nucleotide mature microRNA sequences in a duplex intermediate. The resultant miRNA duplex then interacts with Argonaute 2, and one strand is incorporated into RISC, while the second RNA strand is degraded.9 The mature microRNA guides binding to complementary sequences, typically found in the 3′ UTRs of target mRNAs, thereby repressing translation and/or degrading the target. The nucleotides at positions 2–7 on the 5’-end of the mature miRNA, referred to as the “seed” region, are important for sequence- based targeting, although other regions of the microRNA sequence can contribute to target recognition. The overall structure of the RISC-microRNA complex also is not well characterized, and this overall structure influences which regions of the sequence are available for target binding, thereby adding further complexity to the interaction.10 1.2 TARGET IDENTIFICATION In terms of microRNA-target interaction, it is well accepted that a single microRNA typically targets tens, and potentially over one hundred, of different mRNA transcripts and that a single mRNA may be targeted by multiple microRNAs at a plurality of mRNA binding sites.11 Considering the number of microRNAs that have been discovered, along with the number of potential messenger RNA targets, the combinatoric complexity between even a small set of microRNAs and their targets is extremely high. Furthermore, the interaction between each RISC-associated microRNA and its target site is complex and not fully understood. It is therefore not surprising that the sensitivity and specificity of current sequence-based prediction methods are not sufficient to provide a useful 2 characterization of targetome interaction networks. For example, although two well- established prediction tools, TargetScan and miRanda-mirSVR, provide specificities (ability to correctly identify non-targets) in the high nineties, their sensitivities (ability to correctly identify true targets) are low, at .52 and .62 respectively.12 In light of this insufficiency with sequence based target prediction, methods that do not rely on sequence homology to predict targets are still needed to advance research into the targetome network of microRNAs, as discussed below. On the other hand, there are a number of laboratory techniques for definitively determining specific microRNA-target interactions, which include genome editing of predicted binding sites; reporter gene assays; gene-expression after miRNA modulation; degradome sequencing; cross-linked immuno-precipitation; and biotin-linked chromatography.13 Although such methods are good at determining whether a microRNA binds a specific target, they are unable